Microsoft Research announced on August 20 that Skala is now available in CP2K and is being integrated into Psi4, FHI-aims, ORCA, and VASP, expanding access to its deep-learning density functional theory system for computational chemistry. The announcement also introduced Skala-1.1, an updated exchange-correlation functional (a rule for approximating electron interactions) trained on 2.5x more data than the first public Skala version, with higher accuracy in thermochemistry, reaction kinetics, and molecular structure prediction.

Skala-1.1 achieved a weighted average error of 2.8 kcal/mol on GMTKN55, a benchmark suite with 55 categories covering thermochemistry, reaction barriers, and noncovalent interactions. Microsoft Research said Skala-1.1 ranked first in 32 of the 55 GMTKN55 categories and outperformed the best, most expensive global hybrid functionals while keeping the computational cost of a meta-GGA functional. The model also provides accurate electron densities, dipole moments, and molecular geometries, according to the article.

The accuracy gains came from expansions to the Microsoft Research Accurate Chemistry Collection, or MSR-ACC, a large set of high-accuracy quantum-chemistry reference data generated with expensive wavefunction methods. Microsoft Research said Skala-1.1 added new data categories, including electron affinities and noncovalent clusters, increasing both the size and diversity of training data. The company described Skala as a continuously improving functional, where each release is designed to replace the previous one as new data, model architectures, and training strategies become available.

Microsoft Research said Skala was first made available through an open-source community release built on (GPU4) PySCF and integrated with ASE, giving researchers access to optimized CPU and GPU performance. The new CP2K integration was completed with the team of Prof. Thomas D. Kühne at the Center for Advanced Systems Understanding, or CASUS. CP2K has more than 25 years of development and is used for DFT simulations, especially large-scale systems and long-timescale molecular dynamics.

Microsoft Research said Skala is also being actively integrated into Psi4, which would make it available in three widely used open-source quantum chemistry packages when combined with CP2K and the PySCF-based Skala Community Edition. Work is also underway with developers of FHI-aims, ORCA, and VASP to broaden access across major computational chemistry and materials science software platforms.

The CP2K implementation was tested against the PySCF implementation of Skala-1.1 using closely matched numerical settings. Microsoft Research said the two implementations agreed to within 0.1 kcal/mol MAD across a representative subset of GMTKN55, with one outlier tied to a challenging radical system. A joint paper with the CASUS team describes the CP2K implementation through GauXC, plus the validation strategy and testing infrastructure.

Microsoft Research also published a benchmarking harness and a living performance report for Skala. The report will track performance across tasks and hardware platforms as new Skala releases, GauXC library improvements, and hardware-specific optimizations become available. The article said Skala can currently deliver performance comparable to semi-local meta-GGAs on CPUs, with overhead disappearing for molecules with more than 20-30 atoms, and on GPUs; Figure 3 says Skala 1.1 has the same GPU cost as r2SCAN, while B3LYP and M06-2X become more expensive for systems with more than ~1000 orbitals and CPU overhead disappears for systems with more than ~300 orbitals.

Microsoft Research announced on August 20 that Skala is now available in CP2K and is being integrated into Psi4, FHI-aims, ORCA, and VASP, expanding access to its deep-learning density functional theory system for computational chemistry. The announcement also introduced Skala-1.1, an updated exchange-correlation functional (a rule for approximating electron interactions) trained on 2.5x more data than the first public Skala version, with higher accuracy in thermochemistry, reaction kinetics, and molecular structure prediction.

Skala-1.1 achieved a weighted average error of 2.8 kcal/mol on GMTKN55, a benchmark suite with 55 categories covering thermochemistry, reaction barriers, and noncovalent interactions. Microsoft Research said Skala-1.1 ranked first in 32 of the 55 GMTKN55 categories and outperformed the best, most expensive global hybrid functionals while keeping the computational cost of a meta-GGA functional. The model also provides accurate electron densities, dipole moments, and molecular geometries, according to the article.

The accuracy gains came from expansions to the Microsoft Research Accurate Chemistry Collection, or MSR-ACC, a large set of high-accuracy quantum-chemistry reference data generated with expensive wavefunction methods. Microsoft Research said Skala-1.1 added new data categories, including electron affinities and noncovalent clusters, increasing both the size and diversity of training data. The company described Skala as a continuously improving functional, where each release is designed to replace the previous one as new data, model architectures, and training strategies become available.

Microsoft Research said Skala was first made available through an open-source community release built on (GPU4) PySCF and integrated with ASE, giving researchers access to optimized CPU and GPU performance. The new CP2K integration was completed with the team of Prof. Thomas D. Kühne at the Center for Advanced Systems Understanding, or CASUS. CP2K has more than 25 years of development and is used for DFT simulations, especially large-scale systems and long-timescale molecular dynamics.

Microsoft Research said Skala is also being actively integrated into Psi4, which would make it available in three widely used open-source quantum chemistry packages when combined with CP2K and the PySCF-based Skala Community Edition. Work is also underway with developers of FHI-aims, ORCA, and VASP to broaden access across major computational chemistry and materials science software platforms.

The CP2K implementation was tested against the PySCF implementation of Skala-1.1 using closely matched numerical settings. Microsoft Research said the two implementations agreed to within 0.1 kcal/mol MAD across a representative subset of GMTKN55, with one outlier tied to a challenging radical system. A joint paper with the CASUS team describes the CP2K implementation through GauXC, plus the validation strategy and testing infrastructure.

Microsoft Research also published a benchmarking harness and a living performance report for Skala. The report will track performance across tasks and hardware platforms as new Skala releases, GauXC library improvements, and hardware-specific optimizations become available. The article said Skala can currently deliver performance comparable to semi-local meta-GGAs on CPUs, with overhead disappearing for molecules with more than 20-30 atoms, and on GPUs; Figure 3 says Skala 1.1 has the same GPU cost as r2SCAN, while B3LYP and M06-2X become more expensive for systems with more than ~1000 orbitals and CPU overhead disappears for systems with more than ~300 orbitals.